Scholarlink

<p>With the growing complexity of water supply systems and the increasing demand for smart management, multi-target water demand forecasting (MTF) has become critical for system-level decision-making, scheduling optimization, and resource allocation. Compared with univariate forecasting tasks, MTF confronts challenges stemming from complex spatiotemporal dependencies, missing data and anomalies, and uncertainties regarding the selection of auxiliary forecasting features. Traditional forecasting approaches, largely developed for single-target settings, struggle to capture inter-target dependencies and spatial heterogeneity effectively. In recent years, Graph Neural Networks (GNNs) have attracted increasing attention due to their ability to model relational structures explicitly. By leveraging either predefined or self-learned inter-target dependencies, GNNs simultaneously capture temporal dynamics and spatial interactions among multiple demand nodes, enabling more comprehensive, accurate, and robust forecasting. The study conducts a systematic comparison of representative GNNs for MTF tasks and benchmarks them against sequence-based baselines, including Long Short-Term Memory and Transformer-based networks. Experimental results show that two representative GNNs, namely MTGNN and MTGODE, achieve enhanced accuracy and stability in multi-step MTF tasks with <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\({R}^{2}&gt;\:0.92\)</EquationSource> </InlineEquation> and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(MAPE&lt;15{\%}\)</EquationSource> </InlineEquation>. The SHapley Additive exPlanations-based feature analysis helps to prioritize auxiliary features and yields improved forecasting accuracy, reaching <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\({R}^{2}&gt;0.94\)</EquationSource> </InlineEquation>. Moreover, GNNs show inherent robustness under data irregularities, with only limited degradation in accuracy and reduced error accumulation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-Target Water Demand Forecasting with Graph Neural Networks: A Comparative Study

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Yacan Man,
  • Xiao Zhou,
  • Rui Yuan,
  • Kuizu Su,
  • Shuming Liu

摘要

With the growing complexity of water supply systems and the increasing demand for smart management, multi-target water demand forecasting (MTF) has become critical for system-level decision-making, scheduling optimization, and resource allocation. Compared with univariate forecasting tasks, MTF confronts challenges stemming from complex spatiotemporal dependencies, missing data and anomalies, and uncertainties regarding the selection of auxiliary forecasting features. Traditional forecasting approaches, largely developed for single-target settings, struggle to capture inter-target dependencies and spatial heterogeneity effectively. In recent years, Graph Neural Networks (GNNs) have attracted increasing attention due to their ability to model relational structures explicitly. By leveraging either predefined or self-learned inter-target dependencies, GNNs simultaneously capture temporal dynamics and spatial interactions among multiple demand nodes, enabling more comprehensive, accurate, and robust forecasting. The study conducts a systematic comparison of representative GNNs for MTF tasks and benchmarks them against sequence-based baselines, including Long Short-Term Memory and Transformer-based networks. Experimental results show that two representative GNNs, namely MTGNN and MTGODE, achieve enhanced accuracy and stability in multi-step MTF tasks with \({R}^{2}>\:0.92\) and \(MAPE<15{\%}\) . The SHapley Additive exPlanations-based feature analysis helps to prioritize auxiliary features and yields improved forecasting accuracy, reaching \({R}^{2}>0.94\) . Moreover, GNNs show inherent robustness under data irregularities, with only limited degradation in accuracy and reduced error accumulation.